JPWO2012093451A1 - Speech recognition system, speech recognition method, and speech recognition program - Google Patents

Abstract

The speech recognition system generates a hypothesis that is a chain of words to be searched as a recognition result candidate for input speech data, and a hypothesis search unit that searches for an optimal solution, and a hypothesis search unit that is searching When it is determined that the word or word string included in the hypothesis of the word or word string is re-phrased and the re-phrase determining means for determining whether or not the word or word string is re-phrased, Transparent word hypothesis generation means for generating a transparent word hypothesis that is a hypothesis in which a word or a word string included in a non-fluent section or a repair section of the rephrasing section including the word or word string is treated as a transparent word.

Description

  The present invention relates to a voice recognition system, a voice recognition method, and a voice recognition program.

  In recent years, the application of speech recognition technology has progressed, and speech recognition technology has been used not only for reading speech from a person to a machine but also for natural speech from a person to a person.

  A rephrasing phenomenon exists as a cause of speech recognition errors. Rephrasing is a phenomenon in which a certain word string is replaced as it is or replaced with another word string and re-uttered.

  In the following, based on the model (Repair Interval Model) described in Non-Patent Document 1, for the section related to rephrasing, three sections, a repair target section, a non-fluidity section, and a repair section (Repair Interval). Divide into sections and assume they are continuous. The restoration target section is a section that is restated by the subsequent utterance. The restoration section is a utterance section in which the preceding utterance section is restated. In addition, the non-fluent section refers to a section in which some sound is uttered after the repair target section and the repair section in order to connect to the subsequent repair section, although the previous utterance section is not rephrased, such as saying and interjection. For example, when there is an input “I like apples, ah, and bananas”, the “apples” part is the restoration target section, the “a” part is the non-fluent section, and the “banana” part is the restoration section. . Note that the restoration target section may be referred to as the previous section. On the other hand, the restoration section may be called a section after rephrasing. Note that the non-fluent section may be included in the section before rephrasing or may be included in the section after rephrasing. Moreover, it may be included in neither and it may be a separate area, or may be omitted. Hereinafter, the section from the repair target section to the repair section may be simply referred to as a restatement section.

  Further, Non-Patent Document 2 describes a language analysis system that analyzes unaffected sentences such as rephrasing in a unified manner. The system described in Non-Patent Document 2 is a system for language analysis using input as text, and is realized as an extension of dependency.

Nakatani, C. and Hirschberg, J, "A speech-first model for repair detection and correction", Proceedings of the 31st annual meeting on Association for Computational Linguistics, 1993, p.46-53 Yasuharu Den, “Analysis of Spoken Language Based on Unified Model”, Natural Language Processing Vol. 4, Num. 1, 1997, p. 23-40

  However, as described in Non-Patent Document 2, in a language analysis system, it is common to analyze while looking at long-distance information as in dependency analysis, but in a speech recognition system, N- It is common to use the gram language model. Therefore, a speech recognition system that uses the N-gram language model cannot see long-distance information, and cannot uniformly analyze speech with ineligibility such as rephrasing.

  Therefore, an object of the present invention is to provide a speech recognition system, a speech recognition method, and a speech recognition program that are robust even when the N-gram language model is used as the language model of the speech recognition system.

  The speech recognition system according to the present invention includes a hypothesis search unit that generates a hypothesis that is a chain of words to be searched as a recognition result candidate for input speech data and searches for an optimal solution, and a hypothesis search unit Calculates the rephrasability of the word or word string included in the hypothesis being searched and rephrased by the rephrase determining means for determining whether or not the word or word string is rephrased, A transparent word hypothesis that is a hypothesis that treats a word or word string included in a non-fluent section or repair section of the rephrasing section including the word or word string as a transparent word when it is determined that there is a word. Word hypothesis generation means, and the hypothesis search means searches for an optimal solution by including the transparent word hypothesis generated by the transparent word hypothesis generation means in the hypothesis to be searched. The features.

  In the speech recognition method according to the present invention, the temporary hypothesis searching means searches the input speech data for an optimal solution while generating a hypothesis that is a chain of words to be searched as a recognition result candidate. In the process, calculate the rephrasability of the word or word string included in the hypothesis being searched, determine whether the word or word string is reworded, and if it is determined to be rephrased, The hypothesis search means searches by generating a transparent word hypothesis that is a hypothesis in which a word or a word string included in a non-fluent section or a repair section of the reword section including the word or word string is treated as a transparent word. An optimal solution is searched by including the generated transparent word hypothesis in the target hypothesis.

  Further, the speech recognition program according to the present invention is a hypothesis search process for searching for an optimal solution while generating a hypothesis that is a chain of words to be searched as a recognition result candidate for input speech data to a computer. In the process of calculating the re-phrase of a word or word string included in the hypothesis being searched and determining whether or not the word or word string is re-phrased, it is determined to be re-phrasing A transparent word hypothesis that generates a transparent word hypothesis that is a hypothesis that treats a word or word string included in a non-fluent section or repair section of the rephrasing section including the word or word string as a transparent word The process is executed, and the hypothesis search process searches for an optimal solution including the generated transparent word hypothesis in the hypothesis to be searched.

  ADVANTAGE OF THE INVENTION According to this invention, even when using an N-gram language model as a language model of a speech recognition system, a speech recognition system, a speech recognition method, and a speech recognition program that are robust in other words can be provided.

It is a block diagram which shows the structural example of the speech recognition system of 1st Embodiment. It is a flowchart which shows an example of operation | movement of the speech recognition system of 1st Embodiment. It is a block diagram which shows the structural example of the speech recognition system of 2nd Embodiment. It is a flowchart which shows an example of operation | movement of the speech recognition system of 2nd Embodiment. It is explanatory drawing which shows an example of the hypothesis before hypothesis generation. It is explanatory drawing which shows the other example of the hypothesis before hypothesis generation. FIG. 10 is an explanatory diagram illustrating an example of a hypothesis generated by regarding a word string of a non-fluent section and a repair section as a transparent word. FIG. 10 is an explanatory diagram illustrating an example of a hypothesis generated by regarding a repair target hypothesis and a word string in a non-fluent section as a transparent word. It is a block diagram which shows the outline | summary of this invention. It is a block diagram which shows the other structural example of the speech recognition system by this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

Embodiment 1. FIG.
FIG. 1 is a block diagram illustrating a configuration example of a speech recognition system according to a first embodiment of this invention. The voice recognition system shown in FIG. 1 includes a voice input unit 1, a voice recognition unit 2, and a result output unit 3. The speech recognition unit 2 includes a hypothesis search unit 21, a determination unit 22, and a hypothesis generation unit 23.

  The voice input unit 1 takes in the voice of the speaker as voice data. The audio data is captured as, for example, an audio feature amount series. The voice recognition unit 2 performs voice recognition using the voice data captured by the voice input unit 1 as an input, and outputs a voice recognition result. The result output unit 3 outputs a speech recognition result.

  The hypothesis search unit 21 calculates the likelihood of the hypothesis, develops hypotheses connected to phonemes and words connected to each hypothesis, and searches for solutions.

  The determination unit 2 assumes a repair target section, a non-fluent section, and a repair section in the word chain of each hypothesis, obtains rephrasability under the assumptions, and determines rephrasability above a threshold.

  The hypothesis generation unit 23 generates a hypothesis in which the words in the word string in the non-fluent section and the repair section are treated as transparent words.

  Regarding rephrasing, acoustic information such as the presence or absence of silent intervals, power, pitch, and presence or absence of sudden changes in speech speed, the type of words in non-fluent sections, the proximity of words in the repair target section and the repair section, etc. It is possible to calculate using the index. These indices may be used singly or may be combined linearly or nonlinearly.

  In the present embodiment, the voice input unit 1 is realized by a voice input device such as a microphone, for example. The voice recognition unit 2 (including the hypothesis search unit 21, the determination unit 22, and the hypothesis generation unit 23) is realized by an information processing apparatus that operates according to a program such as a CPU, for example. The result output unit 3 is realized by, for example, an information processing device that operates according to a program such as a CPU and an output device such as a monitor.

  Next, the operation of this embodiment will be described. FIG. 2 is a flowchart showing an example of the operation of the speech recognition system of the present embodiment. In the example shown in FIG. 2, first, the speech recognition unit 1 captures a speaker's utterance as speech data (step A101).

  Next, the voice recognition unit 2 performs voice recognition on the voice data using the fetched voice data as an input. Here, first, the hypothesis search unit 21 calculates the likelihood of an intra-word hypothesis in which the word is not fixed in the input speech data (step A102). Moreover, the hypothesis search part 21 gives a language likelihood based on the confirmed word about the hypothesis which reached | attained the word end (step A103). The intra-word hypothesis is the process of searching for speech data along the time axis from the front, and in the part where the word is uncertain, the word with the same phoneme as one hypothesis It refers to the unit (unit) that is handled. Therefore, in the stage at step A102, the hypothesis search unit 21 performs likelihood calculation in the form of “acoustic likelihood + approximate language likelihood” for the intra-word hypothesis where the word is not fixed. Note that the word likelihood of the word chain is accurately calculated and summed with “acoustic likelihood + language likelihood” when the hypothesis reaches the end of the word and the word is finalized. To do.

  In the process in which the hypothesis search unit 21 searches for hypotheses, the determination unit 22 enumerates a set of repair target sections, non-fluid sections, and repair sections in order from the determined word string. A pair of eyes is taken out (step A104). Here, the determination unit 22 uses the hypothesis generated by the hypothesis search unit 21 (that is, the hypothesis being searched) as one type of word, and sets the predetermined rephrase section setting information. Based on the above, a repair target section, a non-fluent section, and a repair section are assumed. It is assumed that the fixed section includes a confirmed word. The repair target section, the non-fluent section, and the repair section may be, for example, sections of consecutive words, the number of words in the repair target section is L, the number of words in the non-fluent section is M, Assuming that the number of words in the restoration section is a section that allows up to N words, all the combinations that the number of words in each section can take may be listed (L, M, N ≧ 0). Hereinafter, the combination of the repair target section, the non-fluid section and the repair section listed in step A104 may be referred to as a hypothetical reword section group, and a section connecting them may be referred to as a hypothetical reword section.

  Next, the determination unit 22 calculates the rephrasing likelihood for the hypothetical rephrasing section set extracted in step A104 (step A105). The rephrasing is acoustic information such as whether there is a silent section or whether there is a sudden change in power, pitch, speech speed, the type of words in the non-fluent section, the closeness of the words in the repair target section and the repair section, etc. It is possible to calculate using the index.

  Next, the determination unit 22 determines whether or not the calculated rephrase likelihood is greater than or equal to a threshold value (step A106). When the rephrasing probability is equal to or greater than the threshold (Yes in Step A106), the hypothesis generation unit 23 generates a hypothesis using the non-fluidic section and the repaired section in the rephrasing section set of the hypothesis as a transparent word (Step A107). ). Here, the transparent word refers to a word that is treated as non-linguistic in the speech recognition process. Therefore, in the case of a transparent word, when calculating the hypothesis language likelihood, the word is removed and the likelihood is calculated. More specifically, the hypothesis search unit 21 calculates the language likelihood of the hypothesis using the N-gram language model on the assumption that there is no such word as a transparent word.

  On the other hand, if the rephrasing likelihood is less than the threshold value (No in step A106), the process proceeds to step A108. In step A108, the determination unit 22 confirms whether there is a set that has not yet been processed in the listed restatement section assumptions. If it remains, the determination unit 22 returns to step A105 and takes out one set from the remaining sets (Yes in step A108). On the other hand, if the processing from step A105 to step A107 is completed for all of the listed assumption rephrasing section sets (No in step A108), the process proceeds to step A109.

  In step A109, the hypothesis search unit 21 determines whether or not the hypothesis search has been completed up to the end of the speech. Here, if the hypothesis search has not ended up to the end of the speech (No in Step A109), the hypothesis search unit 21 returns to Step A102 and is determined to add or restate the hypothesis generated in Step A107. Then, the hypothesis search for the next speech frame is performed (the processing from steps A102 to A108 is performed on the next speech frame).

  On the other hand, when the hypothesis search has been completed up to the end of the speech (Yes in step A109), the result output unit 3 uses the N-gram language model as a speech recognition result with the hypothesis finally having the maximum likelihood. Output (step A110).

  As described above, in the present embodiment, a rephrasing interval is sequentially assumed for the hypothesis under search, the rephrasing likelihood is calculated, and the non-fluidity interval and the restoration of the interval determined to be rephrased as a result By generating a transparent word hypothesis that dynamically treats a section as a transparent word, it is possible to perform speech recognition with high accuracy with respect to an utterance restoration target section including rephrasing using an N-gram language model.

Embodiment 2. FIG.
Next, a second embodiment of the present invention will be described. FIG. 3 is a block diagram illustrating a configuration example of the speech recognition system according to the second embodiment of this invention. The speech recognition system shown in FIG. 3 is different from the first embodiment shown in FIG. 1 in that the speech recognition unit 2 includes a result generation unit 24.

  In the present embodiment, the hypothesis generation unit 23 generates not only a hypothesis that treats words in the word strings of the non-fluent section and the repair section as transparent words, but also the words in the word string of the repair target section and the non-fluent section. Generate a hypothesis that treats as a transparent word.

  The result generation unit 24 generates a maximum likelihood hypothesis when a hypothesis in which a word string on the repair target section side is treated as a transparent word and a maximum likelihood hypothesis in which a hypothesis is treated with the word string on the repair section side as a transparent word. A speech recognition result combined with a hypothesis is generated.

  Next, the operation of this embodiment will be described. FIG. 4 is a flowchart showing an example of the operation of the voice recognition system of this embodiment. Compared with the operation of the first embodiment, the operation of the present embodiment generates a hypothesis in which the words in the word string in the repair target section and the non-fluent section on the repair target section side are treated as transparent words, or the repair section The system maintains a transparency flag that determines whether to generate a hypothesis that treats words in the non-fluent section and repair section word strings as transparent words, and treats the repair target section as a transparent word The difference is that two maximum likelihood hypotheses are generated: the maximum likelihood hypothesis when the generated hypothesis is generated and the maximum likelihood hypothesis when the hypothesis in which the word string on the repair section side is treated as a transparent word is generated.

  In the example shown in FIG. 4, first, the voice input unit 1 captures the utterance of the speaker as voice data (step A201). In this embodiment, the voice recognition system sets the transparency flag held in the system to the restoration section side at the timing when the voice data is captured (step A202). The transparent flag is information indicating whether a transparent word is created on the repair target section side or on the repair section side.

  Next, the hypothesis search unit 21 calculates the likelihood of a hypothesis in a word whose word is not fixed in the input voice data (step A203). Moreover, the hypothesis search part 21 gives a language likelihood based on the confirmed word about the hypothesis which arrived at the word end (step A204).

  On the other hand, the determination unit 22 enumerates a set of repair target sections, non-fluent sections, and repair sections in order from the determined word string, and extracts the first set (step A205). These sections include fixed words, and the repair target section, the non-fluent section, and the repair section may be, for example, one continuous word, the number of words in the repair target section is L, and the number of words in the non-fluent section is M. The number of words in the repair section may be enumerated as N sections, and all combinations may be listed (L, M, N ≧ 0).

  Next, the determination unit 22 calculates rephrasing for the listed repair target sections, non-fluent sections, and repair sections (step A206). The rephrasing is acoustic information such as whether there is a silent section or whether there is a sudden change in power, pitch, speech speed, the type of words in the non-fluent section, the closeness of the words in the repair target section and the repair section, etc. It is possible to calculate using the index.

  The determination unit 22 determines whether or not the calculated restatement is equal to or greater than a threshold value (step A207). If the rephrasing is equal to or greater than the threshold (Yes in step A207), the hypothesis generation unit 23 sets the repair target section and the non-fluent section as transparent words if the transparency flag held in the system is on the repair target section side. If the transparency flag is on the repair section side, a hypothesis with a non-fluent section and a repair section as transparent words is generated. (Step A208). The hypothesis generated by the hypothesis generation unit 23 is calculated by the hypothesis search unit 2 using the N-gram language model, assuming that there is no transparent word.

  On the other hand, if the rephrasing probability is less than the threshold value, the process proceeds to step A209 (No in step A207).

  In step A209, the determination unit 22 confirms whether the listed combination of the repair target section, the non-fluent section, and the repair section remains. If a set of sections remains (Yes in step A209), the process from step A205 to step A208 is performed for the set of sections.

  On the other hand, if no section set remains (No in step A209), the process proceeds to step A210. In step A210, the hypothesis search unit 21 determines whether or not the hypothesis search has been completed up to the end of the speech. If the hypothesis search has not been completed to the end of the speech (No in step A210), the hypothesis search is completed for the next speech frame. Processing from step A203 to step A209 is performed.

  If the hypothesis search has been completed up to the end of the speech (Yes in Step A210), it is determined whether or not the current transmission flag is on the repair section side (Step A211). (Step A212), and the processing from step A203 to step A210 is similarly performed on the input voice.

  If the current transparency flag is not on the repair section side but on the repair target section side (No in step A211), the result generation unit 24 processes the maximum likelihood hypothesis of the hypothesis on the repair section side that has been processed earlier and the subsequent processing. Compare the hypothesis of the maximum likelihood hypothesis on the repair target section side. Then, the result generation unit 24 determines whether the repair section is selected as a transparent word in the maximum likelihood hypothesis on the repair section, or whether the repair target section is selected as a transparent word in the maximum likelihood hypothesis on the repair target section. Then, a result of combining these two maximum likelihood hypotheses is generated for this rephrasing section (step A213). The result generation unit 24 does not select the repair target section as a transparent word when the repair section is not selected as a transparent word in the maximum likelihood hypothesis on the repair section or when the repair target section is not selected as a transparent word in the maximum likelihood hypothesis on the repair target section. In such a case, it is assumed that these sections are not rephrasing sections, and a result is generated using the maximum likelihood hypothesis based on normal likelihood determination as the maximum likelihood hypothesis of the section without performing this combination process. That is, the result generation unit 24 confirms that a hypothesis having a predetermined interval as a transparent word is selected as the maximum likelihood hypothesis for rephrasing both of the two maximum likelihood hypotheses. Combine the two maximum likelihood hypotheses for the restatement interval.

  The result output unit 3 outputs the result generated by the result generation unit 24 (step A214).

  As described above, in the present embodiment, the transparent word hypothesis in which the repair target section and the non-fluent section are treated as transparent words, and the transparent word hypothesis in which the non-fluent section and the repair section are treated as transparent words are generated. By combining the maximum likelihood hypothesis and outputting the result as a speech recognition result, speech recognition can be performed with high accuracy for the utterance restoration target section including rephrasing even if the N-gram language model is used.

  That is, by generating a transparent word hypothesis in which the repair target section and the non-fluent section are treated as transparent words, the N-gram language model of the word before the repair target section, the repair section, and the word after the repair section is used. be able to. Further, by generating a transparent word hypothesis that treats the non-fluent section and the repair section as transparent words, the N-gram language model of the word before the repair target section, the repair target section, and the word after the repair section is used. be able to. By combining these hypotheses based on the language likelihood of the hypothesis including these two types of transparent words, the word string before the repair target section, the repair target section, the non-fluent section, the repair section, and the repair section While the N-gram language model of the subsequent word string is appropriately adapted, it is possible to output a speech recognition result faithful to the uttered speech.

  In addition, when outputting a hypothesis combining these two hypotheses as a speech recognition result, by adding the information of the reworded section and outputting it, when the output speech recognition result is analyzed by the language analysis system By using the given information, it becomes possible to perform language analysis more accurately.

  In the above description, the example of performing the same process as the repair section side in the process on the repair target section side is shown, but the transparent word hypothesis that treats the word strings of the non-fluent section and the repair section as transparent words is reused. Then, it is also possible to generate only a transparent word hypothesis that treats a word string of a repair target section and a non-fluent section as a transparent word only for a reproducible part.

  In the above description, the transparent word is generated first from the repair section side, but may be generated from the repair target section side. In addition, assuming that the maximum likelihood determination is performed separately, two types of transparent word hypotheses (transparent word hypothesis in which a word string in a non-fluent section and a repair section is treated as a transparent word, and repair are performed in one reword determination. It is also possible to generate a transparent word hypothesis in which the word strings of the target section and the non-fluent section are treated as transparent words.

  Next, a first embodiment of the present invention will be described with reference to the drawings. This example corresponds to the first embodiment. In this example, “Pen, do n’t write in blue” (Japanese utterance: In English shown in FIG. 6, the example is “a pena bed, you know, please write by a blue onea brown one is made of woods” The operation will be described by taking as an example the case of recognizing the utterance ")".

  First, in step A101, the voice input unit 1 reads the speaker's "pen, don, blue so write" (Japanese utterance: English example is "a beda pen, you know, a brown one is made of" Say “woodsplease write by a blue one”).

  Next, in step A102, the hypothesis search unit 21 calculates the likelihood of an intra-word hypothesis that is not determined as to which word is received by using the acquired speech data as an input. This processing is performed, for example, on the utterance of the phoneme of / i / with the word “write” (Japanese: “made of woodsplease write” in the English example) of this utterance example. The acoustic likelihood is calculated with the phoneme model of / and the language likelihood of the previous word chain of the hypothesis such as “blue so” (Japanese: “a brown one is by a blue one” in the English example) It corresponds to adding up.

  Next, in step A103, the hypothesis search unit 21 gives a language likelihood based on the confirmed word for the hypothesis that has reached the end of the word.

  FIG. 5 is an explanatory diagram showing examples of hypotheses searched in this example. In FIG. 5, each ellipse indicates a word (word hypothesis) to be searched as a recognition result candidate. The numerical value attached to each word hypothesis represents the log likelihood of the word chain in which each word hypothesis is linked to the preceding word hypothesis. In this example, when the word “n-” (Japanese: English example is “you know”) is confirmed, the preceding “pen” (Japanese: English example is “a bed”) Is "pen" (Japanese: English example is "a bed"), and "penn" (Japanese: English example is "a beda pen, you know") The language likelihood of word chain is given. In the example illustrated in FIG. 5, a log likelihood of “−60” is given. At the same time, a word chain hypothesis such as “Pan-n” (Japanese: “breada pet, you know” in the English example) may be calculated. In this example, the log likelihood of “−50” is calculated. Is given.

  Next, in step A104, the determination unit 22 enumerates possible combinations of the repair target section, the non-fluent section, and the repair section in the confirmed word string, and takes out the first set. For example, the repair section includes the word determined in step A103, and the repair target section, the non-fluent section, and the repair section may be one continuous word, the repair target section is L words, and the non-fluent section is M words. All the combinations may be listed as continuous intervals allowing up to N words for the repair interval. For example, assuming that the restoration target section is one word, the non-fluent section is one word, and the restoration word is one word, in this utterance example, the word “blue” (Japanese: “a brownblue” in the English example) in step A103. Is fixed, "pen" (Japanese: "a bed" in the English example) and "n-" as the non-fluent section (Japanese: "you know" in the English example) As a repair section, a section set of “blue” (in Japanese: “a brownblue” in the English example) is listed.

  Next, in step A105, the determination unit 22 calculates a rephrase likelihood for the one restatement section set of one hypothesis assumed and extracted in step A104. In the present embodiment, acoustic information such as the length of a silent section, power, pitch, and presence / absence of a sudden change in speech speed is used as an index of restatement. For acoustic information, the length, power, pitch, and speech speed of the silent section are used using learning data that is pre-tagged with the restoration target section, non-fluidity section, and restoration section, and also with acoustic information. Modeling is performed using a mixed Gaussian distribution with time differentiation as a feature amount, and the likelihood with the model is calculated.

  Next, in step A106, the determination unit 22 determines whether or not the restatement probability of the extracted one assumed restatement section is equal to or greater than a threshold value. If the rephrase is greater than or equal to the threshold, the process proceeds to step A107, and if less than the threshold, the process proceeds to step A108.

  In step A107, the hypothesis generation unit 23 generates a hypothesis that regards a word string in a non-fluent section and a repair section as a transparent word for a hypothesis having a rewordability equal to or greater than a threshold value. Remove the deemed word and recalculate the likelihood. Note that the recalculation of the language likelihood of the generated hypothesis may be executed by the hypothesis search unit 21.

  In this utterance example, the non-fluent section is “n-” (Japanese: “you know” in the English example), the repair section is “blue” (Japanese: “a brownblue” in the English example) and “ FIG. 7 shows an example of a hypothesis generated when it is assumed that “No” (Japanese: English example is “one”). In the example shown in FIG. 7, based on the hypothesis shown in FIG. 5, the non-fluent section “n-” (Japanese: “you know” in the English example) and the repair section “blue” (Japanese: In the English example, it is understood that a new hypothesis having “a brownblue”) and “no” (Japanese: “one” in the English example) is newly generated is generated. For this hypothesis, the word “n-” (in Japanese: English example, “you know”) and the repaired segment “blue” (Japanese: English example) Is "a brownblue") and "no" (in Japanese: English example, "one"), except for "write with pen" (Japanese: English example, "please write by a pena language likelihood is given as if it were a chain of words "bed is made of woods"). In this example, the log-likelihood given to the word chain "N-blue so" (Japanese: English example is "you know, a brown one is by a blue one") is "0" (Japanese: English example: “a bed isby a pen”) is given a high log likelihood of “−10”. In this example, the acoustic likelihood is not changed.

  Next, in step A108, the determination unit 22 confirms whether other combinations of the repair target section, the non-fluent section, and the repair section listed in step A104 remain. When it remains, it returns to step A104, takes out one combination from the remaining combinations, and repeats the processing from step A104 to step A107 in the same manner.

  Next, in step A109, the hypothesis search unit 21 determines whether or not the hypothesis search has been completed up to the end of the speech. If the end of the voice has not been reached, the process returns to step A102, and the hypothesis search unit 21 adds the hypothesis generated in step A107 and searches for the hypothesis of the next voice frame. If the end of the voice is reached, the process proceeds to step A110.

  In step A110, the result output unit 3 will finally sound the most likely hypothesis, “write with a pen” (in Japanese: “a bed is made of woodsplease write by a pen”). Output the recognition result.

  By using this example, it was regarded as a non-fluent section and a repair section from the calculated rephrasingness. (Japanese: English example is "you know, a brown oneby a blue one" ) Dynamically as transparent words, the word before the restoration target section and the "pen" (Japanese: English example is "a bed") and the word after the restoration section The distance to "de" (in Japanese: "byis" in the English example) is shortened. Therefore, even with the N-gram language model that is used in conventional speech recognition, it is better to use “pen” than “please write by a bread a pet is made of woods” in Japanese. "Let's write" (Japanese: English example is "a bed is made of woodsplease write by a pen"). As a result, even if the N-gram language model is used, speech recognition can be performed with high accuracy for the utterance restoration target section including rephrasing.

  Next, a second embodiment of the present invention will be described with reference to the drawings. This example corresponds to the second embodiment. In this example, as in the first example, “pen, don, blue so write” (Japanese utterance: in English example, “a beda pen, you know, a brown one is made of woodsplease The operation will be described by taking as an example the case of recognizing the utterance "write by a blue one").

  First, in step A201, the voice input unit 1 reads the speaker's "pen, don, blue so write" (Japanese utterance: English example is "a beda pen, you know, a brown one is made of Say “woodsplease write by a blue one”).

  Next, in step A202, the speech recognition system generates a hypothesis that treats words in the word string of the repair target section and the non-fluent section on the repair target section side as transparent words, or the non-fluent section on the repair section side. And a transparency flag for determining whether to generate a hypothesis in which the words in the word string in the repair section are treated as transparent words are set on the repair section side.

  When this flag is set on the restoration section side, the operation from step A203 to step A210 is the same as the operation from step A102 to step A109 in the first embodiment.

  Next, in step A211, the transparency flag is initially set on the restoration section side, so the process proceeds to step A212, and in step A212, the transparency flag is set on the restoration target section side. In the next steps A203 to A207, the operation is the same as in the first embodiment.

  Next, in step A208, since the transparency flag is a restoration target section, the hypothesis generation unit 23 assumes a word string in the restoration target section and the non-fluent section as a transparent word with respect to a hypothesis having a restatement greater than or equal to a threshold value. Is generated. Then, the hypothesis generation unit 23 recalculates the likelihood by removing words regarded as transparent words in terms of language.

  Figure 8 shows that in this utterance example, the restoration target section is “Pan” (Japanese: “breada pet” in the English example) or “Pen” (Japanese: “a bed” in the English example), and the non-fluent section Is an explanatory diagram showing an example of a hypothesis generated when it is assumed that “n-” (Japanese: English example is “you know”). As shown in FIG. 8, in this example, “pan” (Japanese: “a pet” bread ”in English: English) or“ pen ”(Japanese:“ a bed ”in English: English) ) And "n-" (Japanese: English example, "you know"), and "A brown one is made of woodsplease" Write by a blue one ")) is considered as a word chain, and language likelihood is given. Therefore, from the beginning of the sentence, the word chain “Pan-n” (Japanese: English example is “breada pet, you know”) and from the beginning of the sentence “Pen-n” (Japanese: English example is “pena” log likelihood given to the word chain "bed, you know") is "0", and "-20" for the word chain and the word chain "blue" (Japanese: "a brownblue" in the English example) High log likelihood is given.

  In step A209, as in the first embodiment, it is determined whether there is any other combination. If there are no other combinations, it is determined in step A210 whether the hypothesis search has been completed up to the end of the speech. If the hypothesis search is completed up to the end of the speech, the process proceeds to step A211. In the next step A211, since the transparency flag is on the repair section side, the process proceeds to step A213.

  In step A213, the result generation unit 23 “writes with a pen”, which is the maximum likelihood hypothesis when the transparency flag is on the repair target section side (in Japanese: English example, “a bed is made of woodsplease write by a pen ") and the maximum likelihood hypothesis when the transparency flag is on the restoration section side (in Japanese: English example is" a brown one is made of woodsplease write by a blue one ") A speech recognition result is generated using two maximum likelihood hypotheses.

  Here, the result generation unit 23 firstly selects “pen” (Japanese: English example is “a bed”, which is a word string of a repair target section whose transparency flag is not a transparent word in the maximum likelihood hypothesis on the repair section side. ")" And "n-" (Japanese: English example, "you know"), which is the word string of transparent words in the non-fluent section. Next, the result generation unit 23 transmits the transparent flag “n-” (Japanese: “you know” in the English example)) that is the word string of the transparent word in the non-fluent section of the maximum likelihood hypothesis of the repair target section. The word string “blue” (Japanese: “a brown onea blue one” in the English example) is extracted as a word string of the repair section that is not a word.

  Then, the result generation unit 23 arranges the word strings in the order of the repair target section, the non-fluid section, and the repair section around the common non-fluid section, and arranges the common word strings after the repair section. It produces a speech recognition result that says "Pen-Nen so I write" (Japanese: English example is "a pena bed, you know, a brown one is made of woodsplease write by a blue one"). Here, the word chain in the reworded section indicated by the maximum likelihood hypothesis determined in a series of search processes in which a hypothesis search is performed while generating a transparent word hypothesis that treats the repair target section side as a transparent word, and the repair section side Transparency of a word in the reworded section by combining the word chain in the reworded section indicated by the maximum likelihood hypothesis determined in a series of search processes in which a hypothesis search is performed while generating a transparent word hypothesis treated as a transparent word What is necessary is just to produce | generate the speech-recognition result which shows those chain | linkages in the state which included all instead of a word.

  Finally, in step A214, the result generated in step A213 is output. Here, “Pen-blue is written” (Japanese: English example: “a beda pen, you know, a brown one is made of woodsplease write by a blue one”) is output as the speech recognition result. The

  According to this embodiment, by creating a speech recognition result by combining the transparent word hypothesis that treats the repair target section side as a transparent word and the maximum likelihood hypothesis of the transparent word hypothesis that treats the repair section side as a transparent word, The N-gram language model of the word string before the repair target section, the repair target section, the non-fluent section, the repair section, and the word string after the repair section is appropriately applied. Therefore, it is possible to reduce misrecognition in utterance including rephrasing.

  Also, instead of outputting only this text information as a speech recognition result, the “pen” (Japanese: English example is “a bed”) is the restoration target section, “n-” (Japanese: English example is , "You know") can be output with the information of the non-fluent section and "blue" (Japanese: English example "a brown onea blue one"). If the speech recognition result with the information on the restoration target section, the non-fluent section, and the restoration section is output, for example, when the speech recognition result is analyzed by the language analysis system, the given information is used. It becomes possible to analyze the language more accurately.

  Next, the outline of the present invention will be described. FIG. 9 is a block diagram showing an outline of the present invention. As shown in FIG. 9, the speech recognition apparatus according to the present invention includes a hypothesis search unit 101, a determination unit 102, and a transparent word hypothesis generation unit 103.

  The hypothesis search means 101 (for example, the hypothesis search unit 21) searches the input speech data to generate a hypothesis that is a chain of words to be searched as a recognition result candidate and searches for an optimal solution. Further, the hypothesis searching unit 101 searches for a hypothesis to be searched including a transparent word hypothesis generated by a transparent word hypothesis generating unit 103 described later.

  The rephrase determination unit 102 (for example, the determination unit 22) calculates the rephrase likelihood of the word or word string included in the hypothesis being searched by the hypothesis search unit 101, and whether the word or word string is reworded. Determine whether or not.

  The transparent word hypothesis generation unit 103 (for example, the hypothesis generation unit 23), when the rephrase determination unit 102 determines that the word is rephrased, the non-fluent section or the non-fluent section of the reword section including the word or word string A transparent word hypothesis that is a hypothesis in which a word or a word string included in the repair section is treated as a transparent word is generated.

  In addition, the rephrasing determination unit 102 performs, for a word or word string included in the hypothesis being searched by the hypothesis searching unit 101, a repair target section including the word or word string in the repair section, a non-fluent section, and a repair section. By assuming a combination, calculating the rephrasing for each combination of the assumed repair target section, the non-fluent section, and the repairing section, and determining whether the calculated restatement is equal to or greater than a predetermined threshold, It is determined whether or not the combination is rephrased, and the transparent word hypothesis generation unit 103 determines whether the word included in the non-fluent section or the repair section of the combination determined to be rephrased by the rephrase determination unit 102 A hypothesis that treats the word string as a transparent word may be generated.

  Further, the transparent word hypothesis generation means 103 uses the repair target section side transparent word hypothesis that treats words or word strings included in the repair target section or non-fluent section as transparent words, and the non-fluent section or repair section as the transparent word hypothesis. The hypothesis search means 101 generates a repair word side transparent word hypothesis in which a word or a word string included in the word is treated as a transparent word, and the hypothesis search means 101 sets the search target hypothesis to the repair target section side generated by the transparent word hypothesis generation means. You may search for an optimal solution including the transparent word hypothesis and the repair section side transparent word hypothesis.

  FIG. 10 is a block diagram showing another configuration example of the speech recognition system according to the present invention. As shown in FIG. 10, the speech recognition system according to the present invention may include a result generation unit 104 (for example, a result generation unit 24) that generates a speech recognition result. In such a case, the hypothesis searching unit 101 includes a first search process for searching for an optimal solution by including the generated repair target section side transparent word hypothesis in the hypothesis to be searched, and the generated repair section side A second search process for searching for an optimal solution by including the transparent word hypothesis in the hypothesis to be searched is performed, and the result generation unit 104 performs the speech recognition result by the first search process and the second search process. A voice recognition result combined with the voice recognition result may be output.

  In addition, the result generation unit 104 relates to the section determined to be rephrased, and the maximum likelihood hypothesis indicated as the speech recognition result by the first search processing is the repair target section side transparent word hypothesis, and the second search When the maximum likelihood hypothesis shown as the speech recognition result by the process is the repair section side transparent word hypothesis, the word chain in the reword section indicated by the repair target section side transparent word hypothesis and the repair section side transparent word hypothesis are The speech recognition result indicating the word chain including all the words in the reworded section without including the transparent words may be output by combining the word chain in the rewritten section shown.

  Although not shown, the speech recognition system according to the present invention includes a result output unit (for example, the result output unit 3) that outputs a speech recognition result, and the result output unit is indicated by a word chain of the maximum likelihood hypothesis. In addition to text information to be read, a speech recognition result to which information on a repair target section, a non-fluent section, or a repair section is added may be output.

  In the speech recognition method according to the present invention, the hypothesis searching means searches the optimum solution while generating a hypothesis that is a chain of words to be searched as a recognition result candidate for the input speech data. When the transparent word hypothesis generation means determines that the word is rephrased, the repair target section side transparent word hypothesis that treats a word or word string included in the repair target section or non-fluent section as a transparent word, A hypothesis that generates a repair word side transparent word hypothesis in which a word or a word string included in the fluency section or the repair section is treated as a transparent word, and the hypothesis search means searches for the generated repair target section side transparent word hypothesis A first search process for searching for an optimal solution included in the first search process and a second search process for searching for an optimal solution by including the generated repair section side transparent word hypothesis in the hypothesis to be searched. Output means, a speech recognition result by the first search processing may output a speech recognition result of a combination of a speech recognition result by the second search processing.

  In addition, the speech recognition program according to the present invention calculates to the computer the rephrasability of the word or word string included in the hypothesis being searched, and restates to determine whether the word or word string is rephrased. First transparent word hypothesis that generates a repair target section side transparent word hypothesis in which a word or a word string included in a repair target section or a non-fluent section is treated as a transparent word when it is determined that the determination process is rewording Generation process, second transparent word hypothesis generation process for generating a repair section side transparent word hypothesis that treats a word or a word string included in a non-fluent section or repair section as a transparent word when it is determined to be rephrasing A first search process for searching for an optimal solution by including the generated repair target section side transparent word hypothesis in a hypothesis to be searched; and the generated repair section side transparent word hypothesis The second search process for searching for an optimal solution included in the hypothesis to be searched, and the voice recognition result obtained by combining the voice recognition result by the first search process and the voice recognition result by the second search process are output. The result output process may be executed.

  Although the present invention has been described with reference to the embodiments and examples, the present invention is not limited to the above embodiments and examples. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  This application claims the priority on the basis of the JP Patent application 2011-002307 for which it applied on January 7, 2011, and takes in those the indications of all here.

  The present invention can be widely used in general speech recognition systems. In particular, the present invention can be suitably applied to a speech recognition system that recognizes speech spoken by people such as lecture speech and dialogue speech.

1 speech input unit 2 speech recognition unit 21 hypothesis search unit 22 determination unit 23 hypothesis generation unit 24 result generation unit 3 result output unit 101 hypothesis search unit 102 determination unit 103 transparent word hypothesis generation unit 104 result generation unit

Next, a first embodiment of the present invention will be described with reference to the drawings. This example corresponds to the first embodiment. In this example, "pen, n, write because of blue" (Japanese utterance: English example shown in Fig. 6 is " a bed, you know, a brown one is made of woods") The operation will be described by taking the case as an example.

First, in step A101, the voice input unit 1 reads the speaker's "pen, don, blue so write" (Japanese utterance: English example is "a bed , you know, a brown one is made of woods " ) Is captured as audio data.

Next, in step A102, the hypothesis search unit 21 calculates the likelihood of an intra-word hypothesis that is not determined as to which word is received by using the acquired speech data as an input. This processing is performed, for example, on the utterance of the phoneme of / i / of the word “write” (Japanese: English example is “made of woods ” ) of this utterance example. perform acoustic likelihood calculation of the phoneme model, "because blue" (Japanese: English example, "a brown one is"), such as corresponding to the sum of the language likelihood of the other party of the word chain of the hypothesis To do.

FIG. 5 is an explanatory diagram showing examples of hypotheses searched in this example. In FIG. 5, each ellipse indicates a word (word hypothesis) to be searched as a recognition result candidate. The numerical value attached to each word hypothesis represents the log likelihood of the word chain in which each word hypothesis is linked to the preceding word hypothesis. In this example, when the word "n-" (Japanese: English example is "you know") is confirmed, the preceding "pen" (Japanese: English example is "a bed") When the word hypothesis is “pen” (Japanese: English example is “a bed”), the word chain of “pen” (Japanese: English example is “a bed , you know”) Language likelihood is given. In the example illustrated in FIG. 5, a log likelihood of “−60” is given. At the same time, a word chain hypothesis such as "Pan-n" (Japanese: English example is " a pet, you know") may be calculated. In this example, a log likelihood of "-50" is given. It has been.

Next, in step A104, the determination unit 22 enumerates possible combinations of the repair target section, the non-fluent section, and the repair section in the confirmed word string, and takes out the first set. For example, the repair section includes the word determined in step A103, and the repair target section, the non-fluent section, and the repair section may be one continuous word, the repair target section is L words, and the non-fluent section is M words. All the combinations may be listed as continuous intervals allowing up to N words for the repair interval. For example, the repair target section one word, one word a non-fluent interval, when one word the repair section, in this utterance example is, "blue" in step A103 (Japanese: English example, "a brown") the word when finalized, the "pen" as a repair target section (Japanese: English example, "a bed"), "Hmm" as a non-fluent interval (Japanese: English example, "you know"), repair "blue" as a section (Japanese: English example, "a brown") that the interval set are listed.

In this utterance example, the non-fluent section is "n-" (Japanese: English is "you know"), the repair section is "blue" (Japanese: English is "a brown " ) and "no" FIG. 7 shows an example of a hypothesis generated when it is assumed that (Japanese: English example is “one”). In the example shown in FIG. 7, based on the hypothesis shown in FIG. 5, the non-fluent section “n-” (Japanese: English is “you know”) and the repair section “blue” (Japanese: English) The example shows that a new hypothesis having transparent words “a brown ” ) and “no” (Japanese: English is “one”) is generated. For this hypothesis, the words were considered transparent word non-fluent section "Hmm" (Japanese: English example, "you know") and the word "blue" in the repair section (Japanese: English example , "A brown " ) and "no" (Japanese: English example is "one"), but "write with a pen" (Japanese: English example is " a bed is made of woods") Language likelihood is given as a word chain. In this example, "I'm over blue" (Japanese: English example, "you know, a brown one is") log-likelihood given to the word chain that is next to "0", "pen" (Japanese : English example is given a high log likelihood of “−10” for the word chain “a bed is ” ). In this example, the acoustic likelihood is not changed.

In step A110, the result output unit 3 outputs a speech recognition result of “write with a pen” (Japanese: English example is “a bed is made of woods ” ), which is a hypothesis that becomes the maximum likelihood finally. .

By using this example, "n-blue" (Japanese: English example is "you know, a brown one " ), which was regarded as a non-fluent section and a repair section, was calculated dynamically By treating them as transparent words, the word before the restoration target section and the “pen” (Japanese: English example is “a bed”) and the word “de” (Japan) Word: The English example has a shorter distance to " i s"). Therefore, even in the N-gram language model used in the conventional speech recognition, “write with bread” rather than “write with bread” (Japanese: English example is “ a pet is made of woods”) (Japan) word: English example, will be able to see the language likelihood plausible is more of "a bed is made of woods" ). As a result, even if the N-gram language model is used, speech recognition can be performed with high accuracy for the utterance restoration target section including rephrasing.

Next, a second embodiment of the present invention will be described with reference to the drawings. This example corresponds to the second embodiment. In this example, as in the first example, “pen, don, blue so write” (Japanese utterance: English example is “a bed , you know, a brown one is made of woods ” ) The operation will be described by taking as an example the case of recognizing the voice.

First, in step A201, the voice input unit 1 uses the speaker's "pen, don, blue so write" (Japanese utterance: English example is "a bed , you know, a brown one is made of woods " ) Is captured as audio data.

8, repair target section is "bread" in the present utterance example (Japanese: English example, "a pet") or "pen" (Japanese: English example, "a bed") is, non-fluent section It is explanatory drawing which shows the example of the hypothesis | generation produced | generated when it is assumed that "n-" (Japanese: English example is "you know"). As shown in FIG. 8, in this example, the "bread" of the repair target section (Japanese: English example, "a pet") or "pen" (Japanese: English example, "a bed") and the non Except for the fluent section "n-" (Japanese: English example is "you know"), the word chain is "blue so write" (Japanese: English example is "a brown one is made of woods " ) Language likelihood is given. Therefore, from the beginning of the sentence, the word chain “Pan-n” (Japanese: English example is “ a pet, you know”) and from the beginning of the sentence “Pen-n” (Japanese: English example is “ a bed, logarithmic likelihood given to the word chain “you know”) is “0”, and the logarithm of “−20” is high for the word chain and the word chain “blue” (Japanese: English example is “a brown ” ) A likelihood is given.

In step A213, the result generation unit 24 “writes with a pen” (Japanese: English example is “a bed is made of woods ” ), which is the maximum likelihood hypothesis when the transparency flag is on the repair target section side. Speech recognition results using two maximum likelihood hypotheses: “Blue so write” which is the maximum likelihood hypothesis when the transparency flag is on the repair section side (Japanese: English example is “a brown one is made of woods ” ) Is generated.

Here, the result generation unit 23 firstly selects “pen” (Japanese: English example is “a bed”, which is a word string of a repair target section whose transparency flag is not a transparent word in the maximum likelihood hypothesis on the repair section side. ") and a word string of the transmitted word of non-fluent section" Hmm "(Japanese: English example," to extract you know "). Next, the result generation unit 23 uses “n-” (Japanese: English example is “you know”) and the transparent word, where the transparent flag is the word string of the transparent word in the non-fluent section of the maximum likelihood hypothesis of the repair target section. “Blue” (Japanese: English example: “a brown one ” ), which is a word string of the repair section that is not defined , is extracted.

Then, the result generation unit 23 arranges the word strings in the order of the repair target section, the non-fluid section, and the repair section around the common non-fluid section, and arranges the common word strings after the repair section. It produces a speech recognition result of “Pen-Nen so I write” (Japanese: English example is “ a bed, you know, a brown one is made of woods ” ). Here, the word chain in the reworded section indicated by the maximum likelihood hypothesis determined in a series of search processes in which a hypothesis search is performed while generating a transparent word hypothesis that treats the repair target section side as a transparent word, and the repair section side Transparency of a word in the reworded section by combining the word chain in the reworded section indicated by the maximum likelihood hypothesis determined in a series of search processes in which a hypothesis search is performed while generating a transparent word hypothesis treated as a transparent word What is necessary is just to produce | generate the speech-recognition result which shows those chain | linkages in the state which included all instead of a word.

Finally, in step A214, the result generated in step A213 is output. In this case, "wrote because Pen'n blue" as a result of speech recognition (Japanese: English example, "a bed, you know, a brown one is made of woods") is output.

In addition, instead of outputting only the text information as the speech recognition result, "pen" (Japanese: English example, "a bed") to repair target section, "Hmm" (Japanese: English example, "you know") to a non-fluent section, "blue one" (Japanese: English example, "a brown one") to grant to the information of the repair period will also possible to output. If the speech recognition result with the information on the restoration target section, the non-fluent section, and the restoration section is output, for example, when the speech recognition result is analyzed by the language analysis system, the given information is used. It becomes possible to analyze the language more accurately.

Claims (10)

A hypothesis search means for generating a hypothesis that is a chain of words to be searched as a recognition result candidate for input speech data, and searching for an optimal solution;
A re-statement determining means for calculating a re-stateability of a word or a word string included in the hypothesis being searched by the hypothesis searching means and determining whether or not the word or word string is re-stated;
A hypothesis that treats a word or word string included in a non-fluent section or repair section of a re-word section including the word or word string as a transparent word when it is determined to be re-worded by the re-word determination means. A transparent word hypothesis generating means for generating a certain transparent word hypothesis,
The hypothesis searching means searches for an optimal solution by including the transparent word hypothesis generated by the transparent word hypothesis generating means in the hypothesis to be searched. The rephrasing determination means assumes a combination of a repair target section, a non-fluent section, and a repair section that include the word or word string in the repair section for the word or word string included in the hypothesis being searched by the hypothesis search means. By calculating the rephrasability for each combination of the assumed repair target section, the non-fluent section, and the repair section, and determining whether the calculated restatement is equal to or greater than a predetermined threshold, Determine if it ’s a rephrase,
The transparent word hypothesis generation means generates a hypothesis that treats a word or a word string included in a non-fluent section or repair section of the combination determined to be restated by the restatement determination means as a transparent word. The speech recognition system described. The transparent word hypothesis generation means includes, as a transparent word hypothesis, a repair target section side transparent word hypothesis that treats a word or a word string included in a repair target section or a non-fluent section as a transparent word, and a non-fluent section or a repair section. Generate a repair word side transparent word hypothesis that treats a word or word string as a transparent word,
The hypothesis search means searches for an optimal solution including the repair target section side transparent word hypothesis and the repair section side transparent word hypothesis generated by the transparent word hypothesis generation means in the hypothesis to be searched. Item 3. The speech recognition system according to Item 2. A result generating means for generating a speech recognition result;
The hypothesis searching means includes a first search process for searching for an optimal solution by including the generated repair target section side transparent word hypothesis in a hypothesis to be searched, and the generated repair section side transparent word hypothesis as a search target. And a second search process for searching for an optimal solution included in the hypothesis to be performed,
The voice recognition system according to claim 3, wherein the result generation unit generates a voice recognition result obtained by combining the voice recognition result obtained by the first search process and the voice recognition result obtained by the second search process. The result generation means relates to the section determined to be rephrased, and the maximum likelihood hypothesis indicated as the speech recognition result by the first search process is a repair target section side transparent word hypothesis, and the second search process When the maximum likelihood hypothesis shown as the speech recognition result by is the repair section side transparent word hypothesis, the word chain in the reword section indicated by the repair target section side transparent word hypothesis and the repair section side transparent word hypothesis indicate The speech recognition system according to claim 4, wherein a speech recognition result indicating a word chain including all words in the rephrasing section without including a transparent word is generated by combining the word chain in the rephrasing section. A result output means for outputting a speech recognition result;
The result output means outputs not only the text information indicated by the word chain of the maximum likelihood hypothesis, but also a speech recognition result to which information on a repair target section, a non-fluent section or a repair section is added. The speech recognition system according to any one of the above. In the process of searching for an optimal solution while the hypothesis search means generates a hypothesis that is a chain of words to be searched as a recognition result candidate for the input speech data,
Calculate the rephrasability of a word or word string included in the hypothesis being searched, determine whether the word or word string is rephrased,
A transparent word hypothesis, which is a hypothesis that treats a word or a word string included in a non-fluent section or a repair section of the rephrasing section including the word or word string as a transparent word when it is determined to be rephrased By generating
The speech recognition method, wherein the hypothesis searching means searches for an optimal solution including the generated transparent word hypothesis in a hypothesis to be searched. In the process of searching for an optimal solution while the hypothesis search means generates a hypothesis that is a chain of words to be searched as a recognition result candidate for the input speech data,
When the transparent word hypothesis generating means determines that the word is rephrased, the repair target section side transparent word hypothesis that treats the word or word string included in the repair target section or the non-fluent section as a transparent word, and the non-fluent section Or by generating a repair word side transparent word hypothesis that treats a word or word string included in the repair section as a transparent word,
The hypothesis search means includes a first search process for searching for an optimal solution by including the generated repair target section side transparent word hypothesis in a hypothesis to be searched, and the generated repair section side transparent word hypothesis as a search target. And a second search process for searching for an optimal solution included in the hypothesis to be performed,
The speech recognition method according to claim 7, wherein the result output means outputs a speech recognition result obtained by combining the speech recognition result by the first search process and the speech recognition result by the second search process. On the computer,
In the process of hypothesis search processing that searches for an optimal solution while generating a hypothesis that is a chain of words to be searched as candidate recognition results for the input speech data,
A re-phrase determination process for calculating the re-phrase of a word or word string included in the hypothesis being searched and determining whether the word or word string is re-phrased;
A transparent word hypothesis, which is a hypothesis that treats a word or a word string included in a non-fluent section or a repair section of the rephrasing section including the word or word string as a transparent word when it is determined to be rephrased Execute the generated transparent word hypothesis generation process,
A speech recognition program for searching for an optimal solution including the generated transparent word hypothesis in a hypothesis to be searched in the hypothesis search process. On the computer,
A re-phrase determination process for calculating the re-phrase of a word or word string included in the hypothesis being searched and determining whether the word or word string is re-phrased;
A first transparent word hypothesis generation process for generating a repair target section side transparent word hypothesis that treats a word or a word string included in a repair target section or a non-fluent section as a transparent word when it is determined to be rephrased;
A second transparent word hypothesis generation process for generating a repaired section side transparent word hypothesis that treats a word or a word string included in a non-fluent section or repaired section as a transparent word when it is determined to be rephrased;
A first search process for searching for an optimal solution by including the generated repair target section side transparent word hypothesis in a hypothesis to be searched;
A second search process for searching for an optimal solution by including the generated repair section side transparent word hypothesis in a hypothesis to be searched; a speech recognition result by the first search process; and a second search process. The speech recognition program according to claim 9, wherein a result output process for outputting a speech recognition result in combination with a speech recognition result is executed.

Images